I love how they show the failure cases: compare that with Gemini 1.5 pro's technical paper that carefully avoids any test where it does not seem like a 100% perf! I think confronting your failures a condition for success, and Google seems much too self-indulgent here.